The high randomness and uncertainty of load data make it difficult to improve the accuracy of short-term load forecasting.In order to improve the accuracy of short-term load forecasting,a short-term load forecasting method based on a combined model of fully adaptive noise ensemble empirical mode decomposition (CEEMDAN),convolutional neural network (CNN) and gated recurrent unit (GRU) was proposed.First,the CEEMDAN model was used to decompose the complex original load sequence into several relatively simple sub-sequences;secondly,CNN and GRU were used to establish a prediction model for each component,and the normalized components were input into the training model to obtain the predictive subsequence.Finally,the results of all components were summarized to get the final prediction results.Experimental results show that the CEEMDAN-CNN-GRU combined model has a significant improvement in accuracy compared with the LSTM model,GRU model,CNN-GRU and CEEMDAN-GRU combined model,with an average increase of 25.08%,23.59%,20.41% and 13.53%.The CEEMDAN-CNN-GRU combined model can extract nonlinear features from historical load data,effectively improve the accuracy of short-term load forecasting,and provide strong support for power system construction.